基于计算机视觉与三维实景模型的高密度城区窗景特征评估方法

Evaluation Method for Window View Characteristics in High-Density Areas Based on Computer Vision and 3D Photo-Realistic Models

关键词:  风景园林  窗户景观  深度学习  三维实景模型  视图可视化

Key words:  landscape architecture  window view  deep learning  3D photorealistic model  view visualization

摘要:

建筑窗景作为“视觉”感知媒介的重要扩展，对居民的公共健康状况有着重要影响。针对既有研究缺少大规模、连续性的街区级窗景视觉评估，提出基于Cesium和Vue的窗景数据采集、处理和分析的全流程框架。以香港中环部分核心区域为研究对象，利用城市三维实景数据，对建筑窗景在不同楼栋、楼层、立面下的场景图像进行识别与抓取。窗景特征界定为尺度、颜色和形状3类可量化的视觉指标，结合DeepLabV3+神经网络模型、色彩丰富度算法、边缘检测算法进行分析，并使用K-means聚类分析总结归纳不同要素配比下的窗景视图。结果显示，提出的方法在香港中环测试中表现较好，其单张图像平均抓取速度为0.14 s，窗景图像的各属性特征识别准确，且主要分为生态景观型、构筑主导型和界面复杂型3类窗景，并实现三维可视化比较分析。研究结果为城市空间中各环境要素的布列与划分提供了新的依据，有助于推动未来城市发展的精准管控与设计。

Abstract:

As an important extension of the medium of visual perception, urban architectural window views have a significant impact on residents' public health. Existing studies lack large-scale, continuous, block-level visual assessments of window views. This study proposed a full-process framework for window view data acquisition, processing, and analysis. The framework was implemented using Cesium and Vue. The workflow integrated automated vantage-point sampling, batch image capture, and standardized metadata recording to ensure reproducibility. This study focused on a core area of Central, Hong Kong, using 3D city model data to identify and capture scene images of building window views across different buildings, floors, and facades. The case area was selected for its high density, diverse morphologies, and strong variability in street-canyon conditions. Sampling covered representative facades and height bands to reflect typical residential and office viewpoints. Camera parameters, view directions, and capture intervals were standardized to reduce procedural variance and support consistent comparison. Window view characteristics were defined as three quantifiable visual metrics: scale, color, and shape. These metrics were analyzed using the DeepLabV3+ neural network, a color-richness algorithm, and an edge-detection algorithm. K-means clustering analysis was then applied to summarize and categorize the types and features of window views under different element ratios. Cluster interpretability was emphasized so that categories correspond to recognizable planning conditions and management levers. The research showed that the proposed method performed well in tests conducted in Central, Hong Kong, achieving an average image-capture speed of 0.14 s per image and accurately identifying the attributes of window view images. The window views were primarily categorized into three types: ecological-landscape, structure-dominated, and complex image-interface, with 3D visualization enabling comparative analysis. Based on the spatial characteristics of the study area, the following measures are recommended. Ecological-landscape window views show higher visual quality. Preserve the integrity and continuity of natural view corridors. Use spatial layout to amplify their positive effects and expand the coverage of highquality views. Structure-dominated and complex image-interface views indicate deficits in natural elements and a strong sense of visual enclosure. Introduce natural elements and improve three-dimensional spatial permeability to enhance the overall visual experience. Employ window view metrics as criteria for evaluation and as binding constraints in planning and design. Specifically, in structuredominated areas, improve the visual permeability of street-canyon spaces and the visibility of natural scenery. Use building setbacks and graduated height controls. In parallel, align with current land use to identify underperforming green spaces or grey areas. Prioritize their conversion into high-visibility landscape units to quickly raise the surrounding window view quality. For complex image-interface areas, undertake facade remediation without compromising safety or function. Reduce cluttered railings, dense grilles, and irregular signage to lower visual noise. Where daylight and ventilation will not be obstructed, apply micro-renewal strategies. For example, add planting at windowsills and balconies, and create rooftop gardens. Increase the proportion of natural elements and the continuity of views. The findings provide a new basis for the composition and spatial distribution of environmental elements in urban spaces, contributing to precise control and design for future urban development. This study is constrained by the research scale and the site-specific characteristics of the study area. Future work could expand the urban sample and include comparative analyses across different city types. In parallel, sampling strategies that incorporate actual building orientation and morphological attributes may improve the fidelity with which irregular facades are represented. Finally, enhancing the fidelity of the digital model with respect to facade details, traffic flows, and vegetation characteristics will strengthen overall accuracy and reliability through higher-resolution data acquisition and more precise analytical methods.